Use the AI Workbench

Before opening the Workbench

Prepare a training source, independent benchmark, and knowledge JSON as shown in Prepare assistant CSV and JSON data. Decide on a new candidate folder that does not already exist.

Stage 1 — Configuration

1. Choose AI Workbench… on the toolbar.
2. Choose Edit settings…. Enter the dataset paths and trusted workspace root.
3. Review Network, Training, Evaluation, Security, Packaging, and Branch Agents. Defaults are conservative; do not lower a gate merely to force a pass.
4. Choose Apply settings.
5. Set Configuration JSON to a durable path, then choose Save JSON.

The stage changes to Passed when the exact settings shown in the Workbench match the saved file.

Stage 2 — Dataset preflight

1. Choose Run preflight.
2. Read the row and intent counts, source digests, warnings, and errors.
3. If it fails, correct the source file or settings, save JSON again, and rerun preflight.

Preflight checks formats, trusted paths, intent coverage, benchmark leakage, knowledge coverage, sensitive text, and resource ceilings. It does not train.

Stage 3 — Train and evaluate

1. Enter a new Candidate directory, such as C:\VFS-AI\candidates\main-001.
2. Choose Train candidate.
3. Watch the status strip for epochs and evaluation. You may request cancellation; wait for acknowledgement before closing.

Training never overwrites an existing candidate folder. A failed or cancelled run does not promote its temporary staging tree.

Stage 4 — Review candidate

1. Choose Inspect candidate.
2. Compare accuracy, macro-F1, minimum per-intent recall, label count, architecture, and the full candidate fingerprint with your release requirements.
3. Investigate a weak intent in the benchmark before changing thresholds or adding data.

Stage 5 — Human approval

1. Enter the accountable person or role in Approval authority.
2. Add a note describing the benchmark and evidence reviewed.
3. Choose Approve candidate… and confirm the displayed fingerprint.

Approval records a separate fingerprint-bound decision. It does not train, modify a VFS, or start a package build.

Stage 6 — Select for packaging

1. Choose Select approved candidate.
2. Save the updated configuration JSON when prompted.
3. Use the command in Build and export a Perl package with --assistant-bundle pointing at that candidate.

If a Chronicle workspace is open, Studio may register the approved candidate as a pending review proposal. Proposal review and package selection remain separate actions.

Resume after closing Studio

Studio restores the last assistant settings and configuration path. Open AI Workbench, verify the displayed JSON path, load it if needed, and inspect an existing candidate. Training still requires a new destination.